Chapter 14 Directional Causation

Chapters 9–11 developed the directional framework for measuring dependence between variables. Co-partial moments were shown to capture asymmetric, nonlinear, and tail-specific joint behavior that classical correlation obscures. The copula interpretation then demonstrated how dependence structure can be separated from marginal distributions entirely.

Dependence alone, however, does not imply causation. Building on the conditional-probability and Bayes machinery from the previous chapter, we now turn to directional influence. Two variables may move together because

  • one variable influences the other,
  • both are driven by a common underlying factor, or
  • the relationship arises from structural constraints in the system.

Identifying the direction and strength of causal influence requires more than a symmetric dependence measure. It requires a framework that can detect which variable is doing the driving.

Classical approaches to this problem rely on Granger causality, which uses parametric time-series regressions to test whether lagged values of one variable improve linear predictions of another. The directional framework offers a different approach: nonlinear lagged relationships are compared in both directions without specifying a parametric regression model. Interpreting that association as a structural causal arrow requires additional identification assumptions.

This chapter develops the directional causation framework in three stages: forming explicitly lagged pairs and placing them on the copula scale, measuring how strongly the lagged cause constrains the effect through partial-moment conditional probability and asymmetric directional dependence, and comparing the two directions in a single normalized index.


14.1 Limitations of Classical Granger Causality

In classical time-series analysis, a variable \(X\) is said to Granger-cause \(Y\) if past values of \(X\) improve prediction of \(Y\) beyond what \(Y\)’s own history provides.

A typical vector autoregressive model takes the form

\[ Y_t = \sum_{i=1}^{p} a_i Y_{t-i} + \sum_{i=1}^{p} b_i X_{t-i} + \varepsilon_t . \]

If the coefficients \(b_i\) are jointly significant, \(X\) is said to Granger-cause \(Y\).

Granger causality asks whether \(X\)’s history improves prediction after accounting for \(Y\)’s own past. The NNS index below instead measures directional lagged association from paired values. It does not condition on the effect’s own history or on third variables. Circularly shifting one series in the time-series null retains its cyclic ordering, but this resampling step is not equivalent to conditioning on own lags. A shared driver with different delays, persistence, or coupled dynamics can make the stronger lagged association point in a different direction from the structural forcing. Own-history or third-variable conditioning would be an extension of the method.

The parametric Granger approach carries four limitations that are familiar from earlier chapters.

Linear specification. The causal relationship between variables is constrained to be linear. Nonlinear effects can be invisible to the regression coefficients even when the causal relationship is strong and unambiguous: for an independent, identically distributed, symmetric zero-centered driver with finite fourth moment and nonconstant \(X^2\), a pure lagged quadratic response has zero population linear slope on that driver. (A linear test misses less when the driver is not centred: for \(X\) uniform on \([0,1]\), \(X^2\) has correlation 0.97 with \(X\).)

Symmetric aggregation. Regression models aggregate deviations symmetrically around the mean. Asymmetric causal effects — where large upward movements in \(X\) drive movements in \(Y\) but small movements do not — are absorbed into the residual.

Distributional assumptions. Inference relies on parametric assumptions about the error distribution.

Model dependence. Results are sensitive to lag length selection, variable inclusion, and specification choices that the investigator must make before seeing the data.

The directional construction avoids a specified linear response model and parametric error distribution. It still requires choices of lag, partition resolution, and null resampling, and it does not remove the identification or specification questions above. Granger-style predictability can also be studied with nonlinear models; it is not inherently limited to the displayed linear VAR.


14.2 Design Goals and Identification Limits

Three design goals motivate the directional index. They specify the intended comparison, rather than a theorem that the statistic identifies every structural causal mechanism.

Goal 1 — Self-comparison neutrality. No variable should be identified as causing itself. When \(X = Y\), the causal measure should return a symmetric result indicating no net directional influence, and the diagonal of the causation matrix should be zero.

Goal 2 — Sensitivity to nonlinear association. The measure should respond to nonlinear relationships, including nonmonotone functions. A finite partition need not resolve every functional relationship, and a detected association does not by itself identify its generating causal arrow.

Goal 3 — Directional comparison. The index should summarize the degree of functional asymmetry between the two variables. When \(X\) strongly determines \(Y\) but \(Y\) only weakly constrains \(X\), the statistic should reflect this imbalance clearly and in a stable, interpretable way.

These goals motivate a measure built from two components: a conditional probability that captures how strongly the position of one variable constrains the distribution of the other, and an asymmetric dependence measure that captures whether that constraint is stronger in one direction than the other. Their combination measures resolved directional asymmetry; it does not establish causal identifiability or universal detection.

For an invertible \(f\), \(X=f^{-1}(Y)\) is also a function of \(Y\), so contemporaneous functional determination alone does not identify a causal direction. Finite partitions can resolve the two directions differently. Time ordering through \(\tau>0\) supplies additional information, but it is not by itself a guarantee of causal identification. A non-invertible \(f\), such as \(Y = X^2\), produces an asymmetric association even without a lag, but that asymmetry belongs to the joint distribution, not to the causal arrow: \(X \sim N(0,1)\) with \(Y = X^2\), and \(Y \sim \chi^2_1\) with \(X = S\sqrt{Y}\) for an independent random sign \(S\), have the same joint law, so no statistic of \((X, Y)\) can tell them apart, and NNS.caus(x, y, tau = 0) reads \(X \to Y\) for both. Without a lag the index measures directional association.


14.3 Lagged Co-Partial Moments

The partial-moment framework extends naturally to temporal relationships. This extension provides the conceptual foundation for the lag-normalization step that follows.

Let \(\{X_t\}\) and \(\{Y_t\}\) be two time series with benchmarks \(t_X\) and \(t_Y\). For a lag \(\tau \ge 0\), define the lagged co-partial moments

\[ \text{CoLPM}_{r,s}^{(\tau)}(X,Y) = E\!\Bigl[(t_X - X_{t-\tau})_+^r\,(t_Y - Y_t)_+^s\Bigr] \]

\[ \text{CoUPM}_{r,s}^{(\tau)}(X,Y) = E\!\Bigl[(X_{t-\tau} - t_X)_+^r\,(Y_t - t_Y)_+^s\Bigr] \]

where \((\cdot)_+ = \max(\cdot,0)\) is the positive-part operator from Chapter 2.

When \(\tau = 0\) these reduce exactly to the contemporaneous co-partial moments from Chapter 10. The lagged versions introduce a temporal asymmetry: because the roles of \(X\) and \(Y\) are evaluated at different time points, exchanging \(X\) and \(Y\) does not produce the same quantity:

\[ \text{CoUPM}_{r,s}^{(\tau)}(X,Y) \neq \text{CoUPM}_{r,s}^{(\tau)}(Y,X) \]

in general. This asymmetry is the partial-moment foundation of directional causation. \(\text{CoUPM}_{r,s}^{(\tau)}\) is large when upward deviations of \(X\) at time \(t-\tau\) tend to be followed by upward deviations of \(Y\) at time \(t\). \(\text{CoLPM}_{r,s}^{(\tau)}\) captures the same co-movement in the downward direction.

Like their contemporaneous counterparts from Chapter 10, lagged co-partial moments are estimated directly from sample data by replacing the expectation with an empirical average over the \(n - \tau\) available observation pairs:

\[ \widehat{\text{CoUPM}}_{r,s}^{(\tau)} = \frac{1}{n-\tau} \sum_{t=\tau+1}^{n} (x_{t-\tau} - t_X)_+^r\,(y_t - t_Y)_+^s. \]

These averages converge to population quantities under an appropriate law of large numbers, for example stationarity and ergodicity of the lagged-pair process with finite displayed moments. Arbitrary dependent or nonstationary series do not satisfy that conclusion automatically.

The causation statistic developed below operationalizes this lagged structure directly: it evaluates each direction on its own lagged pairs, places them on the copula scale, and compares the two directions.


14.4 Lagged Pairs on the Copula Scale

Let \(\tau \ge 0\) denote the lag. The two candidate directions are evaluated on two different sets of pairs:

\[ X \to Y:\quad \bigl(X_{t-\tau},\; Y_t\bigr), \qquad Y \to X:\quad \bigl(Y_{t-\tau},\; X_t\bigr), \qquad t = \tau+1, \dots, n . \]

Each direction pairs the earlier value of the candidate cause with the later value of the candidate effect, exactly as in the lagged co-partial moments above. When \(\tau = 0\) both sets are the contemporaneous pairs, which corresponds to the cross-sectional case where temporal ordering carries no information. The lag may be specified directly, or set to \(\tau = \texttt{"ts"}\), which uses one lag for both directions: the mean of the first seasonal period of each series detected by \(\texttt{NNS.seas}\), restricted to at most \(\sqrt{n}\).

Within each set of pairs, both variables are mapped to the copula scale through their own empirical CDFs, the degree-zero lower partial moment from Chapter 3:

\[ U = L_0(X; X) = P(X' \le X) \in (0, 1]. \]

Chapter 12 described copula coordinates; empirical ranks are their sample counterpart, with ties retained. For a fixed lag and alignment, working on ranks makes the statistic invariant to shifts, positive rescaling, and strictly increasing transformations of either variable: adding a constant to \(Y\), changing its units, or replacing it by \(\log Y\) leaves every quantity below unchanged. A decreasing transformation reverses the ranks and can change the index. A data-selected lag, such as tau = "ts", can also change after a transformation, so invariance of the fixed-lag score is not a claim about every lag-selection workflow.

Two features make the comparison directional without removing internal dynamics. First, the lag introduces the temporal asymmetry of the lagged co-partial moments: \((X_{t-\tau}, Y_t)\) and \((Y_{t-\tau}, X_t)\) are different pairs. Second, the comparison is always bidirectional. If the two directional scores are equal, their contributions cancel in the normalized index. A common driver or persistence does not generally guarantee equal scores. In particular, different delays can make a shared cause appear as direction from the series it reaches first.


14.5 Conditional Probability via Partial Moments

With each direction’s pairs on the copula scale, the conditional probability that the position of the cause constrains the distribution of the effect can be measured directly with partial moments.

For the \(X \to Y\) pairs, write \(U_X\) for the copula-scale values of \(X_{t-\tau}\) and \(U_Y\) for those of \(Y_t\). The unconditional deviation of the effect about its mean \(\mu_Y\), over its \(m = n - \tau\) values, is

\[ D(U_Y) = \frac{m}{m-1}\Bigl[L_2(\mu_Y; U_Y) + U_2(\mu_Y; U_Y)\Bigr], \]

the lower plus upper partial moment of degree two — the variance, by Chapter 5. With the factor \(m/(m-1)\) it is a U-statistic, \(\frac{1}{2m(m-1)}\sum_{i \ne j}(U_{Y,i} - U_{Y,j})^2\), and so has the same expectation in any random subset of the sample as in the whole, for any distribution, ties included. Partition the cause into \(k = \lfloor\sqrt{n-\tau}\rfloor\) equal-frequency cells \(c\) on its copula scale, and measure the same deviation of the effect within each cell, about the cell’s own mean:

\[ D_c(U_Y) = D(U_Y \mid c). \]

The deviation is measured quadratically about the cell mean. A split between separated response values therefore retains within-cell variation, as expected when the cause does not resolve the sign. Other dispersion measures, such as the mean absolute pairwise difference, also detect spread but use a different loss geometry.

The quantity called partial-moment conditional probability here is a normalized share of the effect’s deviation resolved by its candidate cause’s cells:

\[ P(X \to Y) = 1 - \frac{\sum_c w_c\, D_c(U_Y)}{D(U_Y)}, \qquad w_c = \frac{n_c}{n-\tau}, \]

bounded to \([0,1]\). This is a resolved-deviation score, not an event probability or a calibrated probability of causation. Under independent random grouping, a nonsingleton cell’s sample variance is unbiased for the global sample variance. Clipping, finite cells, and subsequent score selection still affect the sampling distribution, so the final score need not have exactly zero expectation under independence.

The measure is zero when the effect is spread as widely inside each cell of the cause as it is overall, and it approaches one when the cause’s position pins the effect down. It is not symmetric: \(P(X \to Y) \ne P(Y \to X)\) in general. For a symmetric nondegenerate driver with \(Y=X^2\), knowing \(X\) fixes \(Y\), while knowing \(Y\) leaves the sign of \(X\) unresolved. Sufficiently fine cells can reveal this asymmetry; the finite score depends on the sample and resolution. It uses no kernel bandwidth or parametric response model, but its cell resolution and inferential assumptions still matter.

The rank transformation puts both variables on a common coordinate scale. Direction is then assessed through resolved within-cell variation, not by comparing their numerical ranges. Empirical ranks with ties need not fill the unit interval, and range overlap alone does not identify directional influence.


14.6 Asymmetric Directional Dependence

Conditional probability measures how strongly the cause constrains the effect, but not whether that constraint is stronger in one direction than the other once the dependence structure is taken into account. Two variables may each constrain the other, or one may respond to the other only in extreme regions.

To capture directional alignment, the framework uses the asymmetric dependence measure from Chapter 10 on each direction’s lagged copula pairs: \(\rho_{X \to Y}\) is the dependence of \(U_{Y_t}\) on \(U_{X_{t-\tau}}\) and \(\rho_{Y \to X}\) the dependence of \(U_{X_t}\) on \(U_{Y_{t-\tau}}\). Each is built from the degree-zero and degree-one co-partial moments of its pair within the recursive partition of the candidate cause (split at gravity, Chapter 10), each cell measured against its own permutation null, so it is the lagged co-partial moment structure of this chapter measured as a dependence.

NNS.dep(..., asym = TRUE) reports the strongest level of that partition. The causal comparison instead averages over levels to avoid maximizing over them. This removes that particular selection step, but does not guarantee identical finite-sample null distributions across all margins and serial-dependence structures. Taking the strongest level would favour the variable with more levels to search, so that a continuous series would appear to lead a binary one.

Implementation detail: asym = TRUE turns on asymmetric dependence, but direction is determined by argument order. NNS.dep(x, y, asym = TRUE) partitions on x and measures the dependence of y on x; NNS.dep(y, x, asym = TRUE) measures the reverse.

Define the excess directional dependence of \(Y\) on \(X\) as

\[ \Delta\rho = \rho_{X \to Y} - \rho_{Y \to X}. \]

When \(\Delta\rho > 0\), the lagged movements of \(X\) are more closely tracked by \(Y\) than the reverse — an additional summary of directional lagged association from \(X\) to \(Y\) beyond what the conditional probability alone captures. When \(\Delta\rho \le 0\), this component contributes nothing to the \(X \to Y\) direction.

The use of asymmetric directional dependence here — rather than Pearson correlation — is a direct consequence of Chapter 10. Copula-space variables still exhibit asymmetric tail co-movements that the Chapter 10 framework is designed to detect. Classical symmetric correlation would average away exactly the directional asymmetry that makes the causation statistic informative.


14.7 The Raw Directional Causation Statistic

The two components — conditional probability and asymmetric directional dependence — are combined into a single statistic.

The raw directional causation value from \(X\) to \(Y\) is

\[ \tilde{C}_{X \to Y} = \frac{1}{2} \Bigl[ P(X \to Y) + \max\!\bigl(\Delta\rho, 0\bigr) \Bigr]. \]

The first term rewards how strongly the lagged cause constrains the effect. The second term adds only the positive directional asymmetry in dependence, so reverse-direction dominance is not allowed to inflate the \(X \to Y\) score.

By construction, \(\tilde{C}_{X \to Y}\in[0,1]\): both components are bounded in \([0,1]\), and the outer factor \(1/2\) averages them.


14.8 Bidirectional Normalization

Directional causation should be interpreted comparatively, not in isolation. Define the reverse-direction raw score \(\tilde{C}_{Y \to X}\) by swapping \(X\) and \(Y\) in the same construction, on the pairs \((Y_{t-\tau}, X_t)\).

The final directional causation index from \(X\) to \(Y\) is then normalized as

\[ C_{X \to Y} = \frac{\tilde{C}_{X \to Y} + \kappa}{\tilde{C}_{X \to Y}+\tilde{C}_{Y \to X} + 2\kappa}, \qquad C_{Y \to X}=1-C_{X \to Y}, \qquad \kappa = \frac{1}{\sqrt{n-\tau}}. \]

The neutral pseudo-count \(\kappa\) uses an \((n-\tau)^{-1/2}\) scale as a regularization choice; it is not a universal estimate of the scores’ sampling variation. Without it, two raw scores that are both close to zero — as they are for unrelated variables — would form a ratio dominated by noise and could land anywhere in \([0, 1]\). It shrinks the ratio toward \(0.5\), with a larger effect when both raw scores are small. Independent persistent series can nevertheless have a dispersed index, so distance from \(0.5\) alone is not a significance criterion. It never changes which side of \(0.5\) the index falls on.

Interpretation:

  • \(C_{X \to Y} > 0.5\): stronger evidence for \(X\) leading \(Y\),
  • \(C_{X \to Y} < 0.5\): stronger evidence for \(Y\) leading \(X\),
  • \(C_{X \to Y} \approx 0.5\): weak directional asymmetry.

For several variables, the causation matrix holds \(C_{X_i \to X_j}\) in row \(i\) and column \(j\), so that \(C_{X_i \to X_j} + C_{X_j \to X_i} = 1\) off the diagonal, and its diagonal is zero by the self-comparison convention. A permutation test of the signed statistic \(C_{X \to Y} - C_{Y \to X}\) provides p-values for either direction. For time series the null circularly shifts one series at the lag used for the observed statistic; for cross-sections it shuffles values. A valid randomization interpretation requires invariance under the chosen transformations. Independent exchangeable observations support the latter; cyclic shifts require suitable assumptions about the null process and treatment of boundaries. Trends, random walks, regime changes, or other nonstationarity need not be shift-invariant. Retaining cyclic order is not a general proof of valid time-series p-values.


14.9 Structural Direction and Dynamical Systems

Lagged predictability and structural forcing can disagree. For example, consider coupled maps

\[ X_{t+1}=X_t(3.8-3.8X_t),\qquad Y_{t+1}=Y_t(3.5-3.5Y_t-0.1X_t). \]

The equations include an \(X\)-to-\(Y\) forcing term and no reverse term. Yet the lagged state of \(Y\) can carry substantial information about \(X\), while \(Y\)’s own dynamics dominate its response. A comparison that does not condition on those own dynamics can therefore favor the reverse lagged association. The NNS index should be interpreted as its defined comparison, not a guaranteed recovery of the arrows in such a system.

Structural interpretation requires a stated design or assumptions addressing own-history effects, common drivers, measurement and sampling, and temporal ordering. Adding own-history-conditioned prediction or conditional partial moments would change the method and should be evaluated as an explicit extension.


14.10 Summary

This chapter formalized directional causation as a three-stage construction: explicitly lagged pairs on the copula scale, partial-moment conditional probability and asymmetric directional dependence for each direction, and a bidirectional index of resolved lagged asymmetry.

The resulting statistic is nonparametric and invariant to increasing transformations of either variable, while avoiding the linear-model restrictions of classical Granger-style tests. It measures directional lagged association without conditioning on own history or third variables. Its regularization and shift null do not supply those missing conditions, and causal interpretation requires additional assumptions.

For an applied macroeconomic example using NNS.nowcast, NNS.caus, noise benchmarks, and strength-of-inference summaries, see Causal Inference Amongst Macroeconomic Variables Using NNS.

The next chapter develops distribution comparison methods, providing nonparametric tools for directly comparing distributions without parametric assumptions.

{Generated for: OVVO Labs Registrant Date: 2026-11-14}